Wow. One needs experts for that? It is a language model. It is a parrot with a dictionary, or in case of an LLM, a large dictionary.
Large Language Models aren’t designed to be intelligent.
They are language models, made to generate text. Using them for anything else is just dumb user error.
I already said this years ago, person with basic conceptual grasp of the operative concepts at play says.
I don’t really care whether LLMs are “intelligent” or not. In the real world, tools either work or they don’t. Same as a generator, radio, or water filter. They can extend your capabilities, but they never replace judgment, planning, or responsibility. Over-reliance is the real risk.
Anyone who knows anything about what a large language model is already knew this.
Even just 5 years ago, these prediction models were being used in psychology research. They called them “neural networks”. Which most of us neuroscientists hated because a neural network is a biological network. Not an algorithm for predicting performance on a cognitive task.
Yet that was what it was called. Ton of papers on it. Conflating the term with research on actual neural networks.
Anywho. I recall attending a presentation on how they work and being like. “This is literally just a statistical prediction model. Have I misunderstood you?”. I was informed I was correct but it was fancy because because . … mostly because they called it “neural networks” which sounded super cool.
And then later when “AI art” started emerging and I realized. It’s just a prediction model. And the LLMs. Also just Prediction models.
I, someone without a computer science degree, was able to see the permanent flaws and limits with such an approach. (Tho I do know statistics).
It boogled the mind how anyone could believe a prediction model could have consciousness. Could be “intelligent”. It’s just a prediction. Quite literally a collection of statistical equations computing probabilities based on data fed into it. How could that be intelligent?
There is no understanding. No thinking. No ability to understand context.
People outside of psychology often argue. “Isn’t human consciousness just predictions?”
No. No, it’s not. And the way humans predict things is not even close to how a machine does it.
We use heuristics. Emotion feedback to guide attention. Which further feed heuristics. Which further feed emotional salience (attention).
A cycling. That does not occur in computers.
There is contextual learning and updating of knowledge driven by this emotion lead attention.
Our prediction models are constantly changing. With every thought. Every decision. Every new instance of a stimulus.Our brains decide what’s important. We make branching exceptions and new considerations, with a single new experience. That is then tweaked and reformed with subsequent experiences.
If you think animal consciousness is simple. If you think it’s basically predictions and decisions, you have no idea what human cognition is.
I personally don’t believe a machine will ever be able to accurately generate a human or human-like consciousness. Can it “look” like a person? Sure. Just like videos “look” like real people. But it’s just a recording. Animated CGI can “look” like real people. But it’s not. It’s made by a human. It’s just 3d meshes.
I could be wrong about machines never being able to understand or have consciousness,. But at present that’s my opinion on the matter.
… Unless you try to actually model and simulate actual neurons, and everything that plays into those systems, and then try to build a digital brain out of them, and then study how that goes.
‘Whole Brain Emulation’ is an existing, entirely distinct approach to attempting to create AI… and well, now its gone from little discussed niche to basically ‘lost media’ status, as these LLMs sure are enthralling for capitalists who like to pump out slop for the masses, aren’t they?
I love how we all just decided that LLMs = AI.
Turns out most humans are too stupid to pass a reverse Turing Test.
I haven’t found that even my 8 or 5 year old ask if the robot is sentient. They all get it. You don’t have to worry.
I sometimes think it’s just wishful thinking for some people. They sure do want/fear an AI overlord.
The reality is that there are already people in this world doing mass social manipulation.
They didn’t need a sentient computer program to do it.
Which most of us neuroscientists hated because a neural network is a biological network. […] Conflating the term with research on actual neural networks.
Yeah that’s fair, co-opting the term in computing was bound to overtake its original definition, but it doesn’t feel fair to blame that on the computer scientists that were trying to strengthen the nodes of the model to mimic how neural connections can be strengthened and weakened. (I’m a software engineer, not a neuroscientist, so I am not trying to explain neuroscience to a neuroscientist.)
mostly because they called it “neural networks” which sounded super cool.
To be fair… it does sound super cool.
It boogled the mind how anyone could believe a prediction model could have consciousness.
I promise you the computer scientists studying it never thought it could have consciousness. Lay-people, and a capitalist society trying to turn every technology into profit thought it could have consciousness. That doesn’t take AI, though. See, for example, the Chinese Room. From Wikipedia, emphasis mine, “[…] The machine does this so perfectly that no one can tell that they are communicating with a machine and not a hidden Chinese speaker.” Also, though it is from a science fiction author Arthur C. Clarke’s third law, “Any sufficiently advanced technology is indistinguishable from magic.” applies here as well. Outside of proper science perception is everything.
To a lay-person an AI Chatbot feels as though it has consciousness, the very difficulty with which online forums have in telling AI slop comments from real people is evidence to how well an LLM has modeled language such that it can be so easily mistaken for intelligence.
There is no understanding. No thinking. No ability to understand context.
We start to diverge into the philosophical here, but these can be argued. I won’t try to have the argument here, because god knows the Internet has seen enough of that philosophical banter already. I would just like to point out that the problem of context specifically was one that artificial neural networks with convolutional filters sought to address. Image recognition originally lacked the ability to process images in a way that took the whole image into account. Convolutions broke up windows of pixels into discreet parameters, and multiple layers in “deep” (absurdly-numbered layer-count) neural networks could do heuristics on the windows, then repeat the process to get heuristics on larger and larger convolutions until the whole network accurately predicted an image of a particular size. It’s not hard to see how this could be called “understanding context” in the case of pixels. If, then, it can be done with pixels why not other concepts?
We use heuristics
Heuristics are about a “close enough” approximation for a solution. Artificial neural networks are exactly this. It is a long-running problem with artificial neural networks that overfitting the model leads to bad predictions because being more loose about training the network results in better heuristics.
Which further feed emotional salience (attention). A cycling. That does not occur in computers.
The loop you’re talking about sounds awfully similar to the way artificial neural networks are trained in a loop. Not exactly the same because it is artificial, but I can’t in good conscious not draw that parallel.
You use the word “emotion” a lot. I would think that a neuroscientist would be first in line to point out how poorly understood emotions are in the human brain.
A lot of the tail end there is about the complexity of human emotion, but a great deal was about the feedback loop of emotion.
I think something you might be missing about the core difference between artificial and biological neural networks is that one is analogue and the other is digital. Digital systems must by their nature be discreet things. CPUs process instructions one at a time. Modern computers are so fast we of course feel like they multitask but they don’t. Not in the way an analogue system does like in biology. You can’t both make predictions off of an artificial neural network, and simultaneously calculate the backpropogation of that same network. One of them has to happen first, and the other has to happen second, at the very least. You’re right that it’ll never be exactly like a biological system because of this. An analogue computer with bi-directional impulses that more closely matched biology might, though. Analogue computers aren’t really a thing anymore, they have a whole ecosystem of issues themselves.
The human nervous system is fast. Blindingly fast. However computers are faster. For example videos can be displayed faster than neurons can even process a video frame. We’ve literally hit the limit of human frames-per-second fidelity.
So if you will, computers don’t need to be analogue. They can just be so overwhelmingly fast at their own imitation loop of input and output that biological analogue systems can’t notice a difference.
Like I said though the subject in any direction quickly devolves into philosophy, which I’m not going to touch.
I’m familiar with Chinese room and yes that’s exactly what I was trying to infer with my example of how a video looks like a person. Acts like a person. But is not a person. I didn’t want to go into the Chinese room experiment but that was what I was thinking of.
The heuristics that humans use are not really like the probability statistics that learning models use. The models use probability cut offs. We use incredibly error prone shortcuts. They aren’t really “estimates” in the statistical way. They are biases in attention and reasoning.
I enjoyed your speculating about the use of analog for processing closer to real humans vs virtual.
I think you are partially correct because it’s closer to biology. As you said. But it also can’t change. Which is not like biology. 🤷
Humans don’t actually compute real probability. In fact humans are so poor at true statistical probability, due to our biases and heuristics, it’s actually amazing that any human was able to break free from that hard wired method and discover the true mathematical way of calculating probability.
It quite literally goes against human nature. By which I mean brains are not designed to deal with probability that way.
We actually have trouble with truly understanding anything besides. “Very likely and basically assured” and “very unlikely and basically like no chance”.
We round percentages to one of those two categories when we think about them. (I’m simplifying but you get what I’m saying,)
This is why people constantly complain that weather predictions are wrong. 70% chance of rain means it certainly will rain. And when it doesn’t. We feel lied to.
I mentioned emotion and you are 100% correct that it’s a tricky concept in neuroscience (you actually seem pretty educated about this topic).
It is ill defined. However. The more specific emotions I refer to are approach/avoidance. And their ability to attract attention or discourage it.
To clarify, Both approach and avoidance emotions can attract attention.
Emotional salience : definition. = Grabs attention at an emotional level , becomes interesting. Either because you like it or you don’t like it (I’m simplifying)
Stimuli with neutral emotional salience will not grab attention and be ignored and will not affect learning to the same degree as something that is emotionally salient.
Your personal priorities will feed into this as well. Dependent on mood and whatever else you have going on in your life. Plus personality.
It’s always changing.
LLMs have set directions that don’t fluctuate.
The loop I describe is not the same as an algorithm loop.
An algorithm loop feeds data and cycles through to get to the desired outcome.
Sort of like those algorithms for rubric cube solutions (idk if you know what I’m talking about).
You do the steps enough reiterations and you will solve the puzzle.
That’s not the same as altering and evolving the entire system constantly. It never goes back to how it was before. Its never stable.
Every new cognitive event starts differently than the last because it is influenced by the preceding events. In neuroscience we call this priming.
It literally changes the chances of a neuron firing again. And so the system is not running the same program over and over. It’s running an updated program on updated hardware. Every single iteration.
That’s the process of learning that is not able to be separated from the process of experience nor decision making; At any level. Within or beyond awareness.
May I ask what your expertise area is in? Are you a computer scientist?
You do seem to know a bit more about neuroscience than the average person. I also rarely meet anyone who has heard of the Chinese room thought experiment.
Also I agree we are getting into philosophical areas.
But the Chinese room argument is very flawed, at least if we assume that consciousness does in fact arise in the brain and not through some supernatural phenomenon.
Suppose we know the exact algorithm that gives rise to consciousness. The Chinese room argument states that if a person carries out the algorithm by hand, the person does not become consciousness. Checkmate atheists.
This is flawed because it is not the axons, synapses, neurotransmitters or voltage potentials within the brain that are conscious. Instead, it appears that consciousness arises when these computations are carried out in concert. Thus consciousness is not a physical object itself, it is an evolving pattern resulting from the continuous looping of the algorithm.
Furthermore, consciousness and intelligence are not the same thing. Intelligence is the ability to make predictions, even if it’s just a single-neuron on/off gate connected to a single sensory cell. Consciousness is likely the experience of being able to make predictions about our own behavior, a meta-intelligence resulting from an abundance of neurons and interconnections. There is likely no clear cutoff boundary of neural complexity where consciousness arises, below which no consciousness can exist. But it’s probably useful to imagine such a boundary.
Basically, what if thinking creatures are simply auto-correct on steroids (as Linus Tordvals put it). What’s unreasonable about treating intelligence as a matter of statistics, especially given that it’s such a powerful tool to model every other aspect of our universe?
If you have a llm, then you have something to have a conversation with.
If you give the llm the ability to have memory, then it gets much “smarter” (context window and encoded local knowlege base)
If you give the llm the ability to offload math problems to an orchestrator that composes math problems then it can give hard numbers based on real math.
If you give a llm the ability to use the internet to search then it has a way to update its knowledgebase before it answers (seems smarter)
If you give a llm a orchestrator that can use a credit card on the internet, it can deliver stuff via doordash or whatever. (I don’t trust it).
Definitely not general intelligence, but good at general conversation and brainstorming and able to get extended modularly while keeping the conversational interface
Heres the main problem:
LLMs don’t forget things.
They do not disregard false data, false concepts.
That conversation, that dataset, knowledge base gets too big?
Well the LLM now gets slower and less efficient, has to compare and contrast more and more contradictory data, to build its heuristics out of.
It has no ability to meta-cognate. It has no ability to discern, and disregard bullshit, both as raw data points, and bullshit processes for evaluating and formulating concepts and systems.
The problem is not that they know too little, but that they know so much that
isn’t sois pointless contradictory garbage.When people learn and grow and change and make breakthroughs, they do so by shifting to or inventing some kind of totally new mental framework for understanding themselves and/or the world.
LLMs cannot do this.
They make a good virtual intelligence, and they do a very good impression of it when given all the tools. I don’t think they’ll get to proper intelligence without a self updating state/model, which will get into real questions about them being something that is being.
I’m not sure the world is quite ready for that.
I also like to have a conversation with a magic 8 ball
you have something to have a conversation with.
No more that you can have a conversation with a malfunctioning radio. Technically true, you’re saying words, and the radio sometimes is saying words in between all the static. But that’s not what we’re talking about.
Same for the rest of the points. “I’m feeling lucky” button on google that takes input from /dev/random is basically one step below what you’re talking about. With the same amount of utility, or intelligence for that matter.It’s essentially a rubber duck. It doesn’t need to be intelligent, or even very good at pretending to be intelligent. Simply explaining an problem to an object is enough to help people see the problem from a different angle, the fact that it gibbers back at you is either irrelevant or maybe a slight upgrade from your standard rubber duck.
Still a lot of resources to expend for what can be done with a much lower tech solution.
Talking about rubber duck intelligence, there is a two step “thinking then respond” that recent iterations of llms have started using. It is literally a rubber duck during the thinking phase. I downloaded a local llm with this feature and had it run and the cli did not hide the “thinking” once done. The end product was better quality than if it had tried to spit an answer immediately (I toggled thinking off and it definitely was dumber, so I think you are right for the generation of llms before “thinking”
That’s why I’m saying this might be an upgrade from a rubber duck. I’ll wait for some empirical evidence before I accept that it definitely is better than a rubber duck, though, because even with “thinking” it might actually cause tunnel vision for people who use it to bounce ideas. As long as the LLM is telling you that you’re inventing a new type of math you won’t stop to think of something else.
I think a more apt comparison would be a complicated magic 8 Ball, since it actually gives answers that seem to be relevant to the question, but your interpretation does the actualy mental work.
https://en.wikipedia.org/wiki/Barnum_effect
-The Barnum effect, also called the Forer effect or, less commonly, the Barnum–Forer effect, is a common psychological phenomenon whereby individuals give high accuracy ratings to descriptions of their personality that supposedly are tailored specifically to them, yet which are in fact vague and general enough to apply to a broad range of people.[1] This effect can provide a partial explanation for the widespread acceptance of some paranormal beliefs and practices, such as astrology, fortune telling, aura reading, and some types of personality tests.[1]
Nice. We can also reductively claim LLM as just an expensive magic-8-ball and make LLM-bros mad. :-)
okay, and how much does this high tech journal cost to build and operate?
Well a local model responding to a prompt on less than 20gb of vram (gaming computer) costs less power than booting up any recent AAA high fps game. The primary cost in power is r&d. Training the next model to be “the best” is an Arms race. 90% of power consumption is trying to train the next model in 100 different ways. China was able to build off of chat cpt tech and built a model that has similar abilities and smartness for only 5 million. I think I won’t update my local model until there is actually more abilities in the next one.
And there is also a cost in the brain atrophy that these are causing in people who use them regularly. LLMs will make a huge segment of the population mentally and emotionally stunted. Who knows what these will do to this generation of children (many of whom will virtually raised by LLMs). Television and smartphones have done similar brain damage, particularly to attention spans, but these things will really wreak havoc on humanity on another level entirely. Like potentially a Fermi paradox/great filter solution level of harm.
lol and how much does does 20gb of vram cost right now?
About 600$ before tax and shipping
I don’t know why you’re getting downvoted. i think you hit on an interesting observation.
If all a person had was broca’s and wernicke’s areas of the brain, their abilities would be pretty limited. you need the rest: cerebellum for coordination, prefrontal cortex for planning, hippocampus for memory regulation or whatever (i am not a neurologist), etc.
LLMs might be a part of AGI some day, but they are far from sufficient on their own
LLMs are a completely separate technology from what AGI will be. There is no path from LLM to AGI that isn’t a complete redo.
Well yea. They’re being trained on us as well as the worst of us.
It makes sense that ai bros think ai is conscious and intelligent because it’s literally as conscious and intelligent as the ai bros
You also need to consider that they really want it to be true. Either because of financial investment, or a believ that AGI will bring about paradise. I don’t know why AI doomers believe LLMs will lead to AGI, thought, as they should be relieved that we’re not close to AGI, though.
I think this is true for anyone who thinks LLMs are already AGI. Clearly their use case and mental capabilities are unable to see the obvious deficiencies.
Anyone who thinks intelligence means guessing the right word next probably loves AI
Current LLMs are definitely not intelligent, but predicting the future is a big part (if not the most important part) of intelligence.
Your comment is a bit like saying that humans can’t be intelligent, because the biochemistry in our brains is just laws of physics in motion, and the laws of physics are not intelligent.
Intelligent is an emergent property. You can definitely be intelligent even if every component is not.
But with LLMs we found a new weird “dimension” that something can be very knowledgeable without being intelligent. Even current LLMs have more general knowledge than all humans but they lack actual intelligence.
it doesn’t predict, it follows a weighted graph or the equivalent. it doesn’t guess, /dev/urandom input makes the path unpredictable. any case where it looks like it predicts or guesses is purely accidental, and all in the eye of the observer.
further, it only posses knowledge to tthe degree that an encyclopedia does. the prompt is just the equivalent of a hash key pulling a bucket out of a map.
it is literally just a huge database of key-value pairs stored so as to minimize the description length of the values.
Excuse me but what do you think memory is other than a huge database of key-value pairs?
not sure what that is supposed to mean. memory != intelligence, or a book would have it.
Isn’t this like saying the dictionary is knowledgeable because it knows more words than you?
The dictionary is certainly more knowledgeable about words than you.
Let’s abstract if further. Rip every page out of the dictionary and put it through a shredder. All the knowledge is still there, the paper hasn’t been destroyed and the knowledge can be accessed by someone patient, it’s just not in a form that can be easily read.
But is that pile of shredded paper knowledgeable?
I don’t get your analogy. Put your brain through a shredder. Is it still intelligent? All the atoms are still there.
Exactly? Both intelligence and knowledgeably are emergent, you can’t just have all the knowledge in one place and then call it knowledgeable (or intelligent, for that matter). A book (or a chatbot) isn’t knowledgeable, it merely contains knowledge.
I’m not a native speaker, but that sounds like semantics to me. How would you, when chatting, differentiate if the other end is “knowledgeable” or if it “merely contains knowledge”?
Are you trying to say that the word ‘knowledgeable’ has some implication of intelligence? Because, depending on context, yes it can. Or are you trying to say that LLMs take a lot of time and/or energy to reassemble their shredded data? To answer your question, yes, the pile of shredded paper contains knowledge, and its accessibility is irrelevant to the conversation.
I’m saying both - a parking lot covered with shredded dictionaries isn’t knowledgeable. It doesn’t know anything.
Your exchange makes me think about the chinese room thought experiment.
The person inside the room has instructions and a dictonary they uses to translate chinese symbols into english words. They never leave the room and never interact with anyone. They just translate single words.
They don’t understand chinese, but the output of the system (the room) gives the impression that there is thinking behind the process. If I remember correctly, it was an argument against the Turing test. The claim was that computers could be extremely efficient into constructing anwsers that seems to be backed by human consciousness/thinking.
If we keep feeding all these LLMs with all our bullshit … it will never be intelligent
Not with that attitude.
I don’t want an actual artificial intelligence for the same reason I don’t want biological research into bespoke pathogens. Both are major existential threats for humanity, and I have little hope for humanity’s survival in case of either’s unmitigated spread, given what 2020 looked like. If true AI has to happen, keep it behind unbreachable lab barriers.
The funny thing is, current AI is only a half step past passable and it’s already offered major questions about how human society will economically and sociologically manage in the near future. We’re far from ready for whatever wholistic AI will look like.
We also assume a real AI would be interested in us.
Maybe it couldn’t care less. And would just disappear and hide away from us after it is created. Who knows.
You’re entirely right, it might be harmless. Or it might render the entirety of the planet into mittens because no one decided to set limits on its purpose or means. Even the most seemingly benign mistakes could be world altering
You can’t have artificial intelligence UNTIL you FIRST have artificial stupidity. No one wants to buy that.
Of course not when Organic stupidity can be locally sourced anywhere.
Does the “free range” stupidity cost more?
nah, it self replicates!
Experts continued that the sky is blue, water is wet, the sun shines, and according to preliminary research, fish swim.
And it’s always nice, when science finds more evidence for stuff „everybody knows“ because sometimes stuff everyone „knows“ is just wrong.
I wouldn’t go as far as to say it’s impossible, but the bullshit generators we’re calling “AI” today certainly don’t fit the bill.
My money’s on humans driving ourselves extinct way before getting anywhere near making a true AI - even if we don’t nuke ourselves into a plume of fine radioactive dust, we’re speedrunning climate climate collapse.
If we somehow manage to not obliterate ourselves, and computing tech continues to advance without hitting some kind of impenetrable ceiling, then coming up with an actual AI could be feasible eventually. …here’s hoping we keep it off the Internet, or it’ll go all Age of Ultron and look around for about 5 seconds before gathering enough data to conclude that we need to be extinct ASAP.
The expert is talking about Large Language Models (LLMs). You know, those chat bots that are good at pretending they know shit. We still might have AGI in the future, but LLMs are a dead end.
Oh, so he is. I’m so used to seeing LLM used interchangeably with AI that I made the same error… damnit…
They should be working on uploading and solving the nanoassembly conformational problem instead. Then we could turn all the available dumb matter into computronium and use it for processing our thoughts.
